Despite the longstanding adage ”an image is worth a thousand words,” generating accurate hyper-detailed image descriptions remains unsolved. Trained on short web-scraped image-text, vision-language models often generate incomplete descriptions with visual inconsistencies. We address this via a novel data-centric approach with ImageInWords (IIW), a carefully designed human-in-the-loop framework for curating hyper-detailed image descriptions. Human evaluations on IIW data show major gains compared to recent datasets (+66%) and GPT-4V (+48%) across comprehensiveness, specificity, hallucinations, and more. We also show that fine-tuning with IIW data improves these metrics by +31% against models trained with prior work, even with only 9k samples. Lastly, we evaluate IIW models with text-to-image generation and vision-language reasoning tasks. Our generated descriptions result in the highest fidelity images, and boost compositional reasoning by up to 6% on ARO, SVO-Probes, and Winoground datasets. We release the IIW-Eval benchmark with human judgement labels, object and image-level annotations from our framework, and existing image caption datasets enriched via IIW-model.

When asked to summarize articles or answer questions given a passage, large language models (LLMs) can hallucinate details and respond with unsubstantiated answers that are inaccurate with respect to the input context. This paper describes a simple approach for detecting such **contextual hallucinations**. We hypothesize that contextual hallucinations are related to the extent to which an LLM attends to information in the provided context versus its own generations. Based on this intuition, we propose a simple hallucination detection model whose input features are given by the ratio of attention weights on the context versus newly generated tokens (for each attention head). We find that a linear classifier based on these _lookback ratio_ features is as effective as a richer detector that utilizes the entire hidden states of an LLM or a text-based entailment model. The lookback ratio-based detector—**Lookback Lens**—is found to transfer across tasks and even models, allowing a detector that is trained on a 7B model to be applied (without retraining) to a larger 13B model. We further apply this detector to mitigate contextual hallucinations, and find that a simple classifier-guided decoding approach is able to reduce the amount of hallucination, for example by 9.6% in the XSum summarization task.

Network pruning has emerged as a potential solution to make LLMs cheaper to deploy. However, existing LLM pruning approachesuniversally rely on the C4 dataset as the calibration data for calculating pruning scores, leaving its optimality unexplored. In this study, we evaluate the choice of calibration data on LLM pruning, across a wide range of datasets that are most commonly used in LLM training and evaluation, including four pertaining datasets as well as three categories of downstream tasks encompassing nine datasets. Each downstream dataset is prompted with In-Context Learning (ICL) and Chain-of-Thought (CoT), respectively. Besides the already intriguingobservation that the choice of calibration data significantly impacts the performance of pruned LLMs, our results also uncover several subtle and often unexpected findings, summarized as follows: (1) C4 is not the optimal choice for LLM pruning, even among commonly used pre-training datasets; (2) arithmetic datasets—when used as calibration data—performs on par or even better than pre-training datasets; (3) pruning with downstream datasets does not necessarily help the corresponding downstream task, compared to pre-training data; (4) ICL is widely beneficial to all data categories, whereas CoT is only useful on certain tasks. Our findings shed light on the importance of carefully selecting calibration data for LLM pruning and pave the way for more efficient deployment of these powerfulmodels in real-world applications. We release our code at: https://github.com/abx393/llm-pruning-calibration-data.

Large language models (LLMs), even when specifically trained to process long input contexts, struggle to capture relevant information located in the middle of their input. This phenomenon has been known as the lost-in-the-middle problem. In this work, we make three contributions. First, we set out to understand the factors that cause this phenomenon. In doing so, we establish a connection between lost-in-the-middle to LLMs’ intrinsic attention bias: LLMs exhibit an U-shaped attention bias where the tokens at the beginning and at the end of its input receive higher attention, regardless of their relevance. Second, we mitigate this positional bias through a calibration mechanism, found-in-the-middle, that allows the model to attend to contexts faithfully according to their relevance, even though when they are in the middle. Third, we show found-in-the-middle not only achieves better performance in locating relevant information within a long context, but also eventually leads to improved retrieval-augmented generation (RAG) performance across various tasks, outperforming existing methods by up to 10 percentage point. These findings open up future directions in understanding LLM attention bias and its potential consequences.

Deploying large language models (LLMs) is challenging because they are memory inefficient and compute-intensive for practical applications. In reaction, researchers train smaller task-specific models by either finetuning with human labels or distilling using LLM-generated labels. However, finetuning and distillation require large amounts of training data to achieve comparable performance to LLMs. We introduce Distilling step-by-step, a new mechanism that (a) trains smaller models that outperform LLMs, and (b) achieves so by leveraging less training data needed by finetuning or distillation. Our method extracts LLM rationales as additional supervision for training small models within a multi-task framework. We present three findings across 4 NLP benchmarks: First, compared to both finetuning and distillation, our mechanism achieves better performance with much fewer labeled/unlabeled training examples. Second, compared to few-shot prompted LLMs, we achieve better performance using substantially smaller model sizes. Third, we reduce both the model size and the amount of data required to outperform LLMs; our finetuned 770M T5 model outperforms the few-shot prompted 540B PaLM model using only 80% of available data on a benchmark, whereas standard finetuning the same T5 model struggles to match even by using 100% of the dataset.

Active learning promises to alleviate the massive data needs of supervised machine learning: it has successfully improved sample efficiency by an order of magnitude on traditional tasks like topic classification and object recognition. However, we uncover a striking contrast to this promise: across 5 models and 4 datasets on the task of visual question answering, a wide variety of active learning approaches fail to outperform random selection. To understand this discrepancy, we profile 8 active learning methods on a per-example basis, and identify the problem as collective outliers – groups of examples that active learning methods prefer to acquire but models fail to learn (e.g., questions that ask about text in images or require external knowledge). Through systematic ablation experiments and qualitative visualizations, we verify that collective outliers are a general phenomenon responsible for degrading pool-based active learning. Notably, we show that active learning sample efficiency increases significantly as the number of collective outliers in the active learning pool decreases. We conclude with a discussion and prescriptive recommendations for mitigating the effects of these outliers in future work.

Datasets extracted from social networks and online forums are often prone to the pitfalls of natural language, namely the presence of unstructured and noisy data. In this work, we seek to enable the collection of high-quality question-answer datasets from social media by proposing a novel task for automated quality analysis and data cleaning: question-answer (QA) plausibility. Given a machine or user-generated question and a crowd-sourced response from a social media user, we determine if the question and response are valid; if so, we identify the answer within the free-form response. We design BERT-based models to perform the QA plausibility task, and we evaluate the ability of our models to generate a clean, usable question-answer dataset. Our highest-performing approach consists of a single-task model which determines the plausibility of the question, followed by a multi-task model which evaluates the plausibility of the response as well as extracts answers (Question Plausibility AUROC=0.75, Response Plausibility AUROC=0.78, Answer Extraction F1=0.665).